Why isnt the R580 a lot faster in Shader intensive games?

[Demirug]

I mean that as you change things to improve a given situation, you might, inadvertently, cause a bug in something that used to work. Or reduce performance in some other way. Unless you regress every single app (at all resolutions and modes) against every single change, problems do crop up. Good news is that we eventually catch them and either revert the change or fix the issue. This is true of all SW, really.

I must confess that I am a little bit surprised. You don’t use an automatic test framework that will check a very large number of real game sequence against every new driver build and compare the results (visual and performance) with older driver versions? I had expected that this is something that the computer at ATI do at night when nobody need them.

 

[Ailuros]

I'm with Reverend - synthetic shader benchmarks have told me essentially nothing in all the reviews I've ever read. If you know the architecture, you'll know more...

I think Reverend meant more something along the line that synthetic shader benchmarks are useless in the wrong hands.

In extension if a reviewer can't interpret results from a synthetic application correctly, then I'm not so sure that his/hers interepretations from real games would be of any actual use too.

All that aside, you'll have an extremely hard time convincing me that either/or IHV are not in fact fine-tuning their drivers for synthetic applications after all. Just as with any key-application/game out there.

Finally you can't make them all happy; let alone both major IHVs. One of the two will always feel unfairly treated in the end.

 

[Jawed]

I must confess that I am a little bit surprised. You don’t use an automatic test framework that will check a very large number of real game sequence against every new driver build and compare the results (visual and performance) with older driver versions? I had expected that this is something that the computer at ATI do at night when nobody need them.

If you've seen pictures of the huge racks upon racks of test PCs that they have to use to test their software...

Jawed

 

[Demirug]

If you've seen pictures of the huge racks upon racks of test PCs that they have to use to test their software...

Jawed

I have seen such pictures. This is one of the reasons why I am surprised. I can not believe that the run all the tests over and over again by hand.

 

[Dave Baumann]

Honestly, the shader compiler group, for UDX, has been around for 2 years and nobody's changed there, that I know of. It's not a trivial group size. The lead is still the same. As for the driver optimization, I can't think of anyone either.

Dave?

As for changes, the shader compiler is continuously improving, and getting better, but that does come at the cost of some regressions.

What I'm getting at is a marked difference in the improvements from R520 to R580 between Rightmark, a PS2.x test, and Shadermark V2.1, which is using PS3.0 shadercode on PS3 boards - likewise, R5xx running PS3.0 is still nowhere near the performance of the ATI PS2.0 path in Splinter Cell: Chaos Theory. Now I haven't looked, but unless Shadermark shaders are just very short in all cases (or shorter than Rightmarks shaders), then we'd probably ideally see similar increases, likewise, unless Splinter Cell's ATI's PS2.0 path is just not doing a lot of work that the 3.0 path is then I would conclude that the optimiser isn't working as efectively with PS3.0 code than is is with PS2.0 (which, given the amount of time you've had to work on PS2.0 in relation to PS3.0, wouldn't necessarily be much of a surprise).

 

[Ailuros]

Considering that I just read your interview with Eric D. and Richard H., this thread is a pleasant followup Uhmmm sorry for the OT.

 

[g__day]

ATi seem to be making a very interesting call here - code your 3d pixel shaders using more maths and fewer texture fetches, if your target is a X1900 card. Another complication for folk dealing with 3d graphics engines.

Reminds me of when 3d Now! and SSE were first introduced. What happens if NVidia take the opposite approach (dumber but more raw power)? A 3d coder now has to ponder major algorithim changes and game design changes to cater for one variant of an IHVs latest video cards.

If NVidia varies its strategy than that could be yet another variant of code paths needed to exploit 3d that has to be fit into your 3d engine.

And you need your shader compilers to be better optimised.

I ponder what 3d game developers think about all this and what there feedback to ATi and NVidia is on such pointed 3d h/w design traits?

 

[_xxx_]

What happens if NVidia take the opposite approach (dumber but more raw power)?

That would be senseless and they know that too..

 

[neliz]

That would be senseless and they know that too..

It's not like Ati is forcing developers to switch their routines right? Ati guestimated the shader/texture usage in the near future and came up with 3:1.
FEAR is a good example of this right? it's shader intensiver than other apps.
It was runing good on r520 in comparison to g70. then nvidia released their magic80 driver and now the r580 shows that "hardware optimized for shader performance" gives serious benefits over hardware that is not.
I doubt that FEAR is optimized for ATi anyway (being twimtbp and all)... r580 shows that r580's setup benefits shader intensive hardware more..

Id looks like the only develloper still interested in upping the textures and texture operations

 

[Razor1]

ATi seem to be making a very interesting call here - code your 3d pixel shaders using more maths and fewer texture fetches, if your target is a X1900 card. Another complication for folk dealing with 3d graphics engines.

Reminds me of when 3d Now! and SSE were first introduced. What happens if NVidia take the opposite approach (dumber but more raw power)? A 3d coder now has to ponder major algorithim changes and game design changes to cater for one variant of an IHVs latest video cards.

If NVidia varies its strategy than that could be yet another variant of code paths needed to exploit 3d that has to be fit into your 3d engine.

And you need your shader compilers to be better optimised.

I ponder what 3d game developers think about all this and what there feedback to ATi and NVidia is on such pointed 3d h/w design traits?

I don't think the direction of GPU architecture will vary, both compaines seem to be going in the same direction. This is seen with the shader tests. Both cards perform at similiar levels at least with current tests.

This kind of shows it self in the pure speed tests in 3dmark 06.

http://www.xbitlabs.com/articles/video/display/radeon-x1900xtx_40.html

There isn't much difference between either the 7800 512 or the x1900xt. Then again they might not be using long enough shaders, and this might give the x1900xt an advantage. But in the next year or so games aren't expected to use long enough shaders to possibly show this.

 

[Moloch]

Well I would like some nice large textures

What was the reason that doom 3/ quake 4 have low res textures and low poly counts?

 

[neliz]

What was the reason that doom 3/ quake 4 have low res textures and low poly counts?

It was tailored for nVidia hardware (rimshot [H] style)

 

[Razor1]

Well I would like some nice large textures

What was the reason that doom 3/ quake 4 have low res textures and low poly counts?

Cost vs. performance vs. target system if Quake 4 used 1024x1024 textures sizes it would need 1.0 gig of vram to have the same number of textures it has now. pretty much other then the characters they use 512x512.

 

[OpenGL guy]

What was the reason that doom 3/ quake 4 have low res textures and low poly counts?

The relatively low polygon counts were due to cost of the shadowing algorithm. Stencil shadows are very expensive and the more polygons you have, the more shadow volumes you need to compute. Basically, you need to extrude all the edges of an object's silouette, so higher polygon counts means more edges.

 

[sireric]

I must confess that I am a little bit surprised. You don’t use an automatic test framework that will check a very large number of real game sequence against every new driver build and compare the results (visual and performance) with older driver versions? I had expected that this is something that the computer at ATI do at night when nobody need them.

Actually, yes, we do. There's racks of systems testing all the time. However, they can miss visual problems and they don't test every single app out there, at all resolutions and modes. That would take weeks or months per runtime. As well, not every single source code change is checked -- Sometimes there are multiple between checks (assuming daily writes). So yes, we do have regressions in performance or quality, at times.

 

[sireric]

Dave?

Nope, he's still around. Not sure if he posts -- "Earth to calling Dave, do you post?"?

What I'm getting at is a marked difference in the improvements from R520 to R580 between Rightmark, a PS2.x test, and Shadermark V2.1, which is using PS3.0 shadercode on PS3 boards - likewise, R5xx running PS3.0 is still nowhere near the performance of the ATI PS2.0 path in Splinter Cell: Chaos Theory. Now I haven't looked, but unless Shadermark shaders are just very short in all cases (or shorter than Rightmarks shaders), then we'd probably ideally see similar increases, likewise, unless Splinter Cell's ATI's PS2.0 path is just not doing a lot of work that the 3.0 path is then I would conclude that the optimiser isn't working as efectively with PS3.0 code than is is with PS2.0 (which, given the amount of time you've had to work on PS2.0 in relation to PS3.0, wouldn't necessarily be much of a surprise).

Well, I don't know about the rightmark/shadermark examples, personally. I know we ran lots of internal testing on shaders, and they ended up doing well. I do know that it will take some more time for the shader compiler to catch up on the 3 ALU thing, as some things can be compiled in more optimal ways for that ratio.

As for splinter cell, from what I remember, it's not doing at all the same work under PS3 as under PS2, so they cannot be compared. I think if you take an X1800 or X1900 and force it to take PS2, it should boost up performance. But I'm speaking from memory here (which I hate doing).

 

[Geo]

I do know that it will take some more time for the shader compiler to catch up on the 3 ALU thing, as some things can be compiled in more optimal ways for that ratio.

Reasonable to expect that X1600 performance benefits from those marginal improvements as well?

 

[sireric]

Reasonable to expect that X1600 performance benefits from those marginal improvements as well?

I think anything that will benefit X1800 will benifit X1600, but there's other stuff that should benefit X1600 that should come up, over time.

 

[Moloch]

Cost vs. performance vs. target system if Quake 4 used 1024x1024 textures sizes it would need 1.0 gig of vram to have the same number of textures it has now. pretty much other then the characters they use 512x512.

The relatively low polygon counts were due to cost of the shadowing algorithm. Stencil shadows are very expensive and the more polygons you have, the more shadow volumes you need to compute. Basically, you need to extrude all the edges of an object's silouette, so higher polygon counts means more edges.

thanks.

 

[Nv500]

To be honest, there are so many games can be labelled as "shader-intensitive", and it happen to be that FEAR is a game that ATI did quite well compare to its competitor, so maybe thats the reason why FEAR become an overplayed shader-intensitive game in a Fanatic-intensitive forum?

IMVHO, AOE 3 is more PS 3.0 shader-intentisive, the reason for that is:

1, In-game setting, even at low res with AA off: Simply switch shader from high Q to ultra-high Q usually you get 50%-70% performance hit.

2, As long as you believe 3DM05 or 06 PS shader tests are shader-intensitive, the performance of those video cards in AOE3 at high Q agrees with 3DM05 or 06 raw PS shader benmark results closely (i.e. G70-512 is 2X faster than R520, and R580 is about as fast as G70-512).